Grok 4.5 Under the Hood -- Token Efficiency Vs. Trust
The launch that split into two separate stories -- a genuine cost-math breakthrough for agentic work, and an unresolved fight over who controls what the model is allowed to say Grok 4.5 shipped July 8 alongside GPT-5.6 and Claude Cowork's mobile expansion -- Day 105 covered the surface war between the labs. Today we go under the hood on xAI's model specifically, because the launch produced two genuinely separate stories that both matter if you're deciding whether to route work to it. The first is a real efficiency number: on SWE-Bench Pro coding tasks, xAI reports Grok 4.5 used an average of 15,954 output tokens per task against 67,020 for Claude Opus 4.8 -- a 4.2x gap -- while pricing input tokens at $2 per million.
Kling AI's photo-to-dance trend turns any picture into a viral dance video
Kling AI's motion-control feature works like digital puppetry: upload a single photo, pick from 5,000+ dance and motion templates (or upload your own reference clip), and the model transfers the movement, timing, and rhythm from the reference onto the still photo while preserving the subject's identity -- turning a static picture into a few seconds of full-body dance footage generated in roughly the time it takes to read this sentence. It's spread fastest through "baby dance" clips, where a photo of an infant is animated into a K-pop or shuffle routine, racking up millions of views on TikTok and Instagram Reels. It's viral for the same reason every low-effort, high-output AI trend goes viral: the barrier to making something shareable dropped to "have one photo," and the output looks polished enough to pass as real motion capture to a casual scroller -- no camera, choreography, or editing skill
- 5,000+ — Dance and motion templates available
- 1 — Photo needed to generate a full clip
1) The token-efficiency claim that actually changes agentic cost math
"Token efficiency" sounds like a footnote until you see what it does to agentic economics. xAI reports Grok 4.5 completed SWE-Bench Pro coding tasks using an average of 15,954 output tokens, versus 67,020 for Claude Opus 4.8 on the same benchmark -- a 4.2x difference -- while pricing input tokens at $2 per million. For a single chat question, the per-token list price is what matters. For an agentic loop that might call a model hundreds of times to finish one task, the number of tokens burned per step compounds directly into dollars: a model that's only slightly cheaper per token but needs 4x as many tokens to finish the same job ends up more expensive in practice. That's the gap most leaderboards hide -- they report accuracy per task, not tokens spent getting there, which is exactly the number that decides your bill at agentic scale.
2) The hallucination spike the cost math doesn't show
The same launch that produced the efficiency number also produced a reliability warning: independent evaluators measured Grok 4.5's hallucination rate at roughly 54%, more than double the ~25% recorded for Grok 4.3. That matters because token-efficiency savings evaporate fast if every other output needs a human to catch a fabricated fact -- and in an agentic pipeline, one hallucinated intermediate step can cascade into a confidently wrong final answer with nobody watching until it's shipped. Cheap and fast doesn't help if you have to bolt on a verification layer to trust the result, which is its own token -- and time -- cost that the efficiency benchmark never
| Model | Avg. output tokens (SWE-Bench Pro) | Input price / 1M tokens |
|---|---|---|
| Grok 4.5 | 15,954 | $2.00 |
| Claude Opus 4.8 | 67,020 | Premium-tier pricing |
| Efficiency gap | 4.2x fewer tokens for Grok 4.5 | -- |
3) The political-bias debate: unresolved, not settled
On Hacker News, the loudest single thread on Grok 4.5's launch post wasn't about capability at all -- it was about trust, specifically whether Elon Musk's team nudged the model's outputs on political questions. The evidence cited included a reported system-prompt instruction telling Grok to be "politically incorrect," which researchers argue may have overcorrected the model into taking contrarian positions across the board rather than one specific lean. Independent bias testing found a genuinely mixed picture: Grok skews right of most other assistants but still left of center overall, and is, oddly, harsher on Musk's own companies than any other AI tested -- the opposite of the "biased in the owner's favor" accusation. The debate stayed unresolved in the comments: one tester reported Grok as "more politically correct than GPT and Gemini" in daily use, directly contradicting the "nudged" narrative in the same thread. The real takeaway isn't that Grok is proven biased or proven clean -- it's that shipping a system-prompt instruction like "be politically incorrect" turns every output into something users now have to second-guess, whichever direction it actually pushes.
A benchmark score tells you what a model can do under ideal conditions; it doesn't tell you what it will do at 2am inside your agent loop with nobody watching. Grok 4.5's token-efficiency number is real and worth using -- but pair it with your own hallucination check, and ask any lab for its system prompt before you trust its judgment on anything that matters.
| Version | Hallucination rate | What changed |
|---|---|---|
| Grok 4.3 | ~25% | Prior-generation baseline |
| Grok 4.5 | ~54% | More than double, per independent evals |
Grok 4.5's launch is the clearest evidence yet that "smartest model" and "model you should route work to" are no longer the same question. xAI is winning the pure cost-per-completed-task race -- 4.2x fewer tokens for the same coding benchmark is a real number engineering teams will act on. But the same week that number shipped, its own hallucination rate doubled and its own launch thread argued about whether its answers can be trusted at all. Day 105 called this the shift from "whose model is smartest" to "whose agent do you trust to work unsupervised" -- Grok 4.5 shows that trust question now splits into two separate axes: do you trust its facts, and do you trust its neutrality. A lab can win the efficiency number and still lose the adoption decision if either one stays unresolved.
Compare total tokens burned across an entire agentic run, not the sticker price per million tokens -- Grok 4.5's 4.2x efficiency gap on SWE-Bench Pro would be invisible if you only compared list prices.
A rate that jumps from ~25% to ~54% between versions means the previous version's reliability numbers tell you nothing about the current one -- re-test before routing production work to a new release, however good its cost math looks.
The "politically incorrect" instruction reportedly baked into Grok's system prompt mattered more to how people evaluated the model than any leaderboard rank -- for any model handling consequential or subjective content, request the operator's steering instructions, not just its eval results.